Author: Peng Zhao (Founder of Zhicifang, Co-Founding Partner of Yunhe Capital) Original by IoT Think Tank
This is my 430th column article.
On August 19, the World Robot Conference will kick off in Yizhuang, Beijing. According to the official announcement, the State-owned Assets Supervision and Administration Commission of the State Council (SASAC) will announce the establishment of the "Central Enterprise Robot Innovation Consortium" at the conference. Its core purpose is condensed into twelve words: joint technology R&D, shared application scenarios, and shared achievements.
If the establishment of the consortium sounds the call for scenario aggregation, then tangible orders with real financial backing have already taken the lead in the market.
In April this year, the State Grid internally issued the "2026 Embodied AI Development Plan," which unveiled an astonishing procurement list: a centralized purchase of approximately 8,500 embodied AI devices throughout the year, with a total investment of about 6.8 billion RMB, focusing on four core scenarios: power grid inspection, live-line working, emergency rescue, and warehousing and logistics.
Among them, the procurement amount for three types of core equipment totals 5.8 billion RMB: 5,000 quadruped inspection robot dogs (corresponding to 1.5 billion RMB), 3,000 dual-arm inspection robots (corresponding to 1.8 billion RMB), and 500 humanoid live-line working robots (corresponding to 2.5 billion RMB). An additional 1 billion RMB is specifically allocated to technology R&D and talent cultivation.
According to the plan, the procurement pace is divided into three batches: pilot in the first quarter, scaling up in the third quarter, and supplementation in the fourth quarter. At this moment, the industry is in a critical window for scaled procurement.
Over the past four months, the entire industry has been excitedly calculating its share of this 6.8 billion RMB order. However, there is a key clause in the document that is easily overlooked: All procured equipment must comply with the technical specifications for power embodied AI devices. Priority will be given to selecting partner manufacturers that can deeply adapt to the Guangming Power Large Model and support localized deployment to ensure data security.
In fact, the order amount is the least scarce thing in this document. What this procurement plan truly delivers to the industry are three core elements that money cannot buy: scenarios, specifications, and cost discipline. What this article aims to do is to break down these three elements one by one, and provide insight into what they really mean for embodied AI, and even the entire AIoT industry.
Value Reconstruction: When the "Brain" Returns to the Demand Side
Let us first answer the most striking question: Why does the buyer bring its own "brain"?
Let us turn the clock back seven years.
In January 2019, the State Grid proposed the construction of the "Ubiquitous Power Internet of Things (IoT)," and the construction outline was released in March of the same year. At that time, the number of terminals connected to the State Grid system had exceeded 500 million (including 450 million smart meters), which invigorated the entire AIoT industry. According to Wind statistics, there were once as many as 43 concept stocks.
However, this carnival lasted only one year. In early 2020, with the change of leadership at the State Grid, the strategic narrative shifted to the "New Power System," causing concept stocks to plummet in response, and the "Ubiquitous Power IoT" gradually faded from the public discourse.
The outside world misread this history as a "receding of concepts," ignoring the most critical undercurrent: although the concept retreated, the construction never stopped.
In May 2019, the State Grid Big Data Center was unveiled, and the IoT management platform, data middle platform, and perception layer access specifications were successively implemented. These achievements quietly settled into the buyer's system, becoming the foundation for all subsequent intelligentization.
If the first act was the buyer building its own connection layer and platform layer, then the second act took place in December 2024.
The State Grid released the hundred-billion-parameter-level multimodal "Guangming Power Large Model" and signed strategic cooperation agreements with Baidu and Alibaba. Here, tech giants have retreated to the role of technology suppliers, while the initiative of the platform remains firmly in the hands of the buyer.
What is even more worth pondering is the technical detail that "Guangming" covers a full range of sizes from the hundred-million-parameter level to the hundred-billion-parameter level, and its edge-side distilled version is already ready within the system. This means that the "brain" running locally on the inspection robot dogs and live-line working robots in the future has already been reserved in advance.
This is by no means an isolated event. From the SASAC deploying the "AI+ Special Action," to China Southern Power Grid's "Da Watt," PetroChina's "Kunlun," and China State Shipbuilding Corporation's "Baige"... large models from central enterprises have already spread like a prairie fire.
And this plan in April this year is precisely the third act: model adaptation has been officially written into the bidding documents, and the decision-making level has completed a closed loop in terms of clauses.
Three acts in seven years; rather than the expansion of buyer power, it is better described as the inevitable return of the "brain" to its proper place.
In high-risk industries such as electric power, every judgment in live-line working is directly tied to grid and personal safety. Once a judgment goes wrong, the ultimate bearer of responsibility is the power grid enterprise.
Placing the "brain" in the hands of the party that holds the full volume of scenario data and bears the ultimate responsibility for accidents is, in essence, the alignment of "decision-making power" and "responsibility power."
Seeing through this layer, the strategic map for suppliers becomes much clearer.
The industry once believed that embodied AI would replicate the script of smartphones: hardware manufacturers competing fiercely on hardware, and tech giants building models to collect an "ecosystem tax." However, vertical industries such as electric power have given a completely different answer: in responsibility-intensive vertical industries, the platform layer should be held by the demand side.
The "intelligence premium" of equipment will give way to "carrier value." Future competition will focus on computing power redundancy, power consumption control, security levels, model deployment efficiency, and full lifecycle management. In-house algorithms are no longer a premium item; deep adaptation capability is.
When a model needs to be deployed to tens of thousands of heterogeneous devices, continuous iteration, grayscale verification, and version rollback will become rigid demands. Just as the connectivity era gave birth to the connectivity management platform, the embodied AI era will inevitably give rise to a platform for the distribution and O&M of judgments.
Price Crucible: Centralized Procurement Mechanism Reshapes the Cost Curve
The second thing is hidden in the budget sheet. We need to first understand three sets of seemingly contradictory data.
Calculated according to the document, the average price of robot dogs is 300,000 RMB, dual-arm inspection robots are 600,000 RMB, and humanoid live-line working robots are 5 million RMB. However, according to calculations by the industrial chain, the current market unit price of inspection robot dogs is between 500,000 and 1 million RMB. At the same time, the document's own economic feasibility argument points to a higher pricing: a single device saves an average of 500,000 to 800,000 RMB in labor costs annually, with an investment payback period of 2 to 3 years. Deducting backwards from this, the equipment value should logically be in the million-RMB range.
The three sets of data seem contradictory, but the implied reasonable explanation is that what is written on the budget sheet is the "terminal price."
The buyer has disclosed in advance to the entire market the cost bottom line that the industry must reach in the next two to three years.
Compared with current market prices, this budget embeds a price reduction requirement of 40% to 70%. It is not just an order, but also a public "cost roadmap."
According to the plan, the penetration rate in key areas will reach 30% in 2026, covering over 80% of high-risk scenarios in 2027, and completing autonomous O&M by 2030. The terminal price and the mileage table are disclosed simultaneously.
In fact, the AIoT industry has a profound muscle memory of this mechanism.
The "three-piece set" of central enterprise centralized procurement: unified technical specifications, multiple companies taking turns winning bids, and inter-batch price linkage moving downward, has indiscriminately baptized hundreds of millions of smart meters and NB-IoT modules. It can compress gross profit to the limit, forcing enterprises' differentiation capabilities to migrate towards chips and platforms.
Seemingly cruel, the development history of smart meters precisely reveals the ultimate trajectory of this curve.
Taking Hexing Electrical, a leading smart meter exporter, as an example, its 2025 annual report shows that the gross profit margin of domestic business dropped by 10 percentage points, while the gross profit margin of overseas business only slightly decreased by 3.4 percentage points. Forging costs domestically and harvesting profits overseas is the best proof of the dual-circulation mechanism.
Looking at it as a whole, the other side of centralized procurement emerges: it is the strictest coach for Chinese enterprises before going global. Global buyers will only pay for costs and reliability verified through large-scale real-world deployment, and these two points can only be forged by centralized procurement in an extremely short time.
Faced with this mechanism, embodied AI enterprises have three cards to play:
1. Going global to play the time difference: Using the ultimate cost tempered by domestic centralized procurement to earn profits in overseas markets where regulations are not yet fully formed, utilizing the window period to export a full set of "integrated software and hardware" solutions.
2. Making big value with the "cerebellum": Since the "brain" has returned to the buyer, suppliers must establish irreplaceability in the "cerebellum" (motion control). Deeply coupling force control, balance, high-frequency real-time response with core components.
3. Extending towards "O&M": The deployment of models by the buyer requires services such as heterogeneous device access and edge-side computing power scheduling. Transitioning from simply "selling equipment" to "selling base operation services" may be the only way forward for hardware manufacturers.
Scenarios: From "Scarce Resources" to "Quota Platforms"
The third thing is the most easily underestimated.
Over the past decade, the biggest dead knot for AIoT startups has never been technology, but real-world scenarios. The standard narrative of the industry is "trading scenarios for validation, and trading validation for financing." For access to a single substation, an enterprise might spend months, only to end up with a mere pilot project. Scenarios are an extremely scarce resource, yet they have never been transformed into quantifiable assets.
Now, let us re-examine the buyer's moves. The State Grid's long-term plan has listed embodied AI as one of the five major intelligent capabilities, covering six major fields and over 600 specific scenarios. This means that within a single buyer, scenarios have undergone comprehensive inventory and cataloging.
And the soon-to-be-established Innovation Consortium will promote this model to the entire central enterprise camp. According to the supporting "Real-scene Practical Training Special Action," each central enterprise is required to open no less than 10 real-world working scenarios, which will be uniformly undertaken by the consortium, sharing data internally, and listing embodied large models as a concentrated direction for tackling key problems.
Inventory, quantification, centralized management, and sharing—when these four words are superimposed, it means that "scenarios" possess the form of quantifiable and distributable assets for the first time in history.
If we place this in a larger macro sequence, the context will become clearer:
2018-2022: Each central enterprise built its own industrial Internet platform (such as Aerospace Cloud Network, CECII), fighting alone;
2023-2025: Each built industry large models (Da Watt, Kunlun, Dayu), still fragmented;
2026: The consortium steps in, uniformly aggregating fragmented scenarios and data.
The demand side has spent eight years thoroughly mastering the "platform approach" and has stepped into the deep water zone that the supply side has not yet touched: transforming the most scarce scenario resources into the industry's infrastructure.
For suppliers, this may signal the retirement of old sales skills. Shifting comprehensively from competing on customer relationships and information asymmetry to competing on contract terms and engineering implementation capabilities.
Under this new set of rules, the compulsory course for suppliers is to realign with the "triple identity" of the buyer: at the hardware body layer, it is the customer of embodied AI enterprises; at the scenario layer, it is the platform holding the access quota; at the model layer, it is the co-creator with overlapping R&D routes.
Concluding Remarks
Scenarios provide the entry ticket, specifications provide the engineering ruler, and cost discipline provides the ticket for going global. The 6.8 billion RMB order is just the most easily replicated number among them.
Industrial history has repeatedly verified this logic. China's photovoltaic and NEV (New Energy Vehicle) industries, without exception, completed the tempering of cost and reliability in the most stringent domestic centralized procurement and extreme competition, and only then seized the throne in the global market.
At the end of "The Count of Monte Cristo," Alexandre Dumas wrote: All human wisdom is contained in these two words, wait and hope.
For the current embodied AI industry, hope lies in three practical paths: going global to harvest profits, using the cerebellum to create value, and taking over the vacancies through O&M.
Waiting may be passive, but hope must be executable. Those who read the order thinly only see the 6.8 billion RMB right in front of them; while those who read the list thickly have already seen the future path of the industry.